hermes-dayshift-glm

human-approved implementer for hermes-nightshift-glm. kanban triage, model lanes, quota-gated GLM execution, Codex execution, and guarded PR repair.

---

dayshift is the companion implementer for hermes-nightshift-glm. Nightshift creates issues and PRs. Dayshift scans those outputs, classifies what is actionable, puts them on a local kanban board, and only runs an executor after a human moves a card into an execution lane.

how it works

1. scan: reads open GitHub issues and PRs that look like Nightshift output
2. classify: scores each item for fixability, risk, effort, and approval path
3. triage: stores local state in ~/.dayshift/state.json and renders a jKanban board
4. approve: human moves cards into model-specific execution lanes
5. execute: runs Codex or Hermes/GLM with the issue or PR context
6. validate: installs lockfile dependencies when needed and runs detected tests/checks
7. output: issue cards create Dayshift PRs; PR cards can be repaired or merged when policy allows

architecture

Dayshift is intentionally conservative:

scan ──→ classify ──→ board ──→ execution lane ──→ agent ──→ validate ──→ PR/merge
                  │                 │                         │
                  └──── local state ┘                         └── failed → retry lane

• local board state is the source of truth: GitHub labels are best-effort sync, but stale labels do not override local moves
• execution lanes are explicit: a lane chooses model, command, run policy, reasoning effort, and whether to merge after execution
• terminal work stays off the board: closed, merged, skipped, ignored, and done records remain in state but do not clutter kanban
• failed means real execution failure: label sync noise is hidden and missing old labels are treated as already-clean

quickstart

requires a GitHub token that can read issues/PRs, apply labels, push branches, and create PRs in the target repos.

mkdir -p ~/dayshift-workspace ~/.dayshift
curl -sL https://raw.githubusercontent.com/Microck/hermes-dayshift-glm/main/dayshift.py > ~/dayshift-workspace/dayshift.py
curl -sL https://raw.githubusercontent.com/Microck/hermes-dayshift-glm/main/scripts/hermes-agent-runner.py > ~/dayshift-workspace/hermes-agent-runner.py
chmod +x ~/dayshift-workspace/dayshift.py ~/dayshift-workspace/hermes-agent-runner.py

save a GitHub token:

printf '%s\n' 'github-token-here' > ~/.dayshift/.gh-token-dayshift
chmod 600 ~/.dayshift/.gh-token-dayshift

scan, run once, or serve the board:

python3 ~/dayshift-workspace/dayshift.py scan
python3 ~/dayshift-workspace/dayshift.py run
python3 ~/dayshift-workspace/dayshift.py serve --host 127.0.0.1 --port 3001

kanban board

The web UI is local-first and intentionally small. It supports:

• model-specific execution columns such as execute: GLM 5.1 and execute: GPT 5.3 Codex
• per-card executor notes, injected into the next agent prompt as human guidance
• local search and bulk selection
• red x close action that closes the GitHub issue or PR
• yellow / ignore action that hides a card locally without closing GitHub
• bulk move, skip, and close actions

Default visible board order:

1. issue inbox
2. PR inbox
3. one execute: <model> column per configured lane
4. ready and merge, shown only when human merge approval is needed
5. in-progress
6. failed, only for real executor or validation errors

done and skip are workflow states, not visible columns. Those records are retained in ~/.dayshift/state.json for auditability.

execution lanes

The default lanes are:

lane	model	runner	policy
dayshift/execute-glm-5-1	glm-5.1	Hermes Agent adapter	waits for GLM quota window
dayshift/execute-gpt-5-3-codex	gpt-5.3-codex	Codex CLI	runs immediately

The GLM lane uses scripts/hermes-agent-runner.py, which passes a prompt file to hermes chat --query. The Codex lane streams the prompt through stdin with:

codex exec --model gpt-5.3-codex -

validation

Dayshift detects common project shapes before accepting agent output:

• Python tests: python3 -m unittest discover
• pnpm repos: pnpm install --frozen-lockfile, then pnpm test
• Bun repos: bun install --frozen-lockfile, then bun test
• Yarn repos: yarn install --frozen-lockfile, then yarn test
• npm repos: npm ci, then npm test
• CMake repos: configure, build, then ctest --output-on-failure

You can override this with validation_commands in config.

config

Dayshift reads ~/.dayshift/config.json. If the file does not exist, defaults are used.

{
  "target_repos": [],
  "repo_discovery_mode": "dayshift",
  "exclude_repos": ["*-backup"],
  "public_only": true,
  "max_inactive_days": 30,
  "min_size_kb": 10,
  "max_repos_to_consider": 30,
  "max_prs_per_repo": 2,
  "github_token_file": "~/.dayshift/.gh-token-dayshift",
  "nightshift_token_file": "~/.nightshift/.gh-token-nightshift",
  "kanban_enabled": true,
  "scheduler_enabled": true,
  "scheduler_interval_seconds": 30,
  "glm_quota_command": "python3 ~/nightshift-workspace/glm_quota.py --check",
  "validation_commands": [],
  "merge_method": "squash"
}

The settings page exposes execution lanes, token file paths, scheduler options, merge policy, repo discovery, and validation commands.

commands

python3 dayshift.py config-path
python3 dayshift.py scan
python3 dayshift.py scan --apply-labels
python3 dayshift.py run
python3 dayshift.py serve --host 127.0.0.1 --port 3001

Environment override:

DAYSHIFT_TARGET_REPOS=Microck/hermes-nightshift-glm python3 dayshift.py scan

files

file	description
dayshift.py	scanner, classifier, scheduler, kanban UI, executor orchestration
scripts/hermes-agent-runner.py	non-interactive Hermes Agent adapter for GLM lanes
test_dayshift.py	unit tests for parsing, state, kanban rendering, scheduler, execution lanes, validation
architecture.md	design notes for scanner, labels, board, executor, and merge policy
classifier.md	classifier heuristics and risk model
input-sources.md	GitHub input source notes
kanban.md	board model and workflow labels
roadmap.md	implementation roadmap

for llm agents

if a user asks you to install or run hermes-dayshift-glm:

1. create directories:

   mkdir -p ~/dayshift-workspace ~/.dayshift

2. download dayshift.py and scripts/hermes-agent-runner.py from this repo

3. ensure a GitHub token exists at ~/.dayshift/.gh-token-dayshift

4. scan a narrow target first:

   DAYSHIFT_TARGET_REPOS=Microck/hermes-nightshift-glm python3 ~/dayshift-workspace/dayshift.py scan

5. start the board:

   python3 ~/dayshift-workspace/dayshift.py serve --host 127.0.0.1 --port 3001

6. move cards into execution lanes only after reviewing title, summary, risk, and approach

Never publish ~/.dayshift/config.json, ~/.dayshift/state.json, GitHub tokens, Hermes config, BrowserStack credentials, or agent runtime logs.

license

mit license